To Build Truly Intelligent Machines, Teach Them Cause and Effect
Judea Pearl, a pioneering figure in artificial intelligence, argues that AI has been stuck in a decades-long rut. His prescription for progress? Teach machines to understand the question why.
Judea Pearl 是人工智能领域的先驱人物，他认为人工智能已经陷入了长达数十年的泥潭。他进步的处方？教机器理解为什么的问题。
READ LATER
Monica Almeida for Quanta Magazine
Kevin Hartnett
Senior Writer/Editor
May 15, 2018
artificial intelligencebig datacomputer sciencemachine learningQ&AAll topics
Artificial intelligence owes a lot of its smarts to Judea Pearl. In the 1980s he led efforts that allowed machines to reason probabilistically. Now he’s one of the field’s sharpest critics. In his latest book, “The Book of Why: The New Science of Cause and Effect,” he argues that artificial intelligence has been handicapped by an incomplete understanding of what intelligence really is.
Three decades ago, a prime challenge in artificial intelligence research was to program machines to associate a potential cause to a set of observable conditions. Pearl figured out how to do that using a scheme called Bayesian networks. Bayesian networks made it practical for machines to say that, given a patient who returned from Africa with a fever and body aches, the most likely explanation was malaria. In 2011 Pearl won the Turing Award, computer science’s highest honor, in large part for this work.
But as Pearl sees it, the field of AI got mired in probabilistic associations. These days, headlines tout the latest breakthroughs in machine learning and neural networks. We read about computers that can master ancient games and drive cars. Pearl is underwhelmed. As he sees it, the state of the art in artificial intelligence today is merely a souped-up version of what machines could already do a generation ago: find hidden regularities in a large set of data. “All the impressive achievements of deep learning amount to just curve fitting,” he said recently.
In his new book, Pearl, now 81, elaborates a vision for how truly intelligent machines would think. The key, he argues, is to replace reasoning by association with causal reasoning. Instead of the mere ability to correlate fever and malaria, machines need the capacity to reason that malaria causes fever. Once this kind of causal framework is in place, it becomes possible for machines to ask counterfactual questions — to inquire how the causal relationships would change given some kind of intervention — which Pearl views as the cornerstone of scientific thought. Pearl also proposes a formal language in which to make this kind of thinking possible — a 21st-century version of the Bayesian framework that allowed machines to think probabilistically.
Pearl expects that causal reasoning could provide machines with human-level intelligence. They’d be able to communicate with humans more effectively and even, he explains, achieve status as moral entities with a capacity for free will — and for evil. Quanta Magazine sat down with Pearl at a recent conference in San Diego and later held a follow-up interview with him by phone. An edited and condensed version of those conversations follows.
工智能的很多智慧都归功于 Judea Pearl。在 1980 年代，他领导了让机器进行概率推理的努力。现在他是该领域最尖锐的批评者之一。在他的最新着作“原因之书：因果关系的新科学”中，他认为人工智能受到了对智能真正含义的不完整理解的阻碍。
三年前，人工智能研究的一个主要挑战是对机器进行编程，以将潜在原因与一组可观察条件相关联。Pearl 使用一种称为贝叶斯网络的方案想出了如何做到这一点。贝叶斯网络使机器可以说，鉴于一名患者从非洲返回发烧和身体疼痛，最有可能的解释是疟疾。2011 年，Pearl 获得了计算机科学最高荣誉图灵奖，这在很大程度上归功于这项工作。
但正如珀尔所见，人工智能领域陷入了概率关联。这些天来，头条新闻都在吹捧机器学习和神经网络的最新突破。我们读到了可以掌握古代游戏和驾驶汽车的计算机。珍珠不知所措。在他看来，当今最先进的人工智能技术只不过是机器在一代人之前已经可以做到的事情的增强版：在大量数据中找到隐藏的规律。“深度学习的所有令人印象深刻的成就都只是曲线拟合，”他最近说。
现年 81 岁的珀尔在他的新书中详细阐述了真正智能机器的思维方式。他认为，关键是用因果推理取代关联推理。机器不仅需要关联发烧和疟疾的能力，还需要推理疟疾导致发烧的能力。一旦这种因果框架到位，机器就可以提出反事实问题——询问因果关系在某种干预下会如何变化——珀尔认为这是科学思想的基石。珀尔还提出了一种形式语言，使这种思考成为可能——21 世纪版本的贝叶斯框架，它允许机器进行概率思考。
Pearl 预计因果推理可以为机器提供人类级别的智能。他们将能够更有效地与人类交流，甚至，他解释说，获得具有自由意志和邪恶能力的道德实体的地位。Quanta Magazine最近在圣地亚哥举行的一次会议上与 Pearl 坐了下来，后来通过电话对他进行了后续采访。这些对话的编辑和浓缩版本如下。
Why is your new book called “The Book of Why”?
It means to be a summary of the work I’ve been doing the past 25 years about cause and effect, what it means in one’s life, its applications, and how we go about coming up with answers to questions that are inherently causal. Oddly, those questions have been abandoned by science. So I’m here to make up for the neglect of science.
为什么你的新书被称为“The Book of Why”？
这意味着要总结过去 25 年来我一直在做的关于因果关系的工作，它在一个人的生活中意味着什么，它的应用，以及我们如何为那些本质上是因果关系的问题提出答案。奇怪的是，这些问题已经被科学抛弃了。所以我来这里是为了弥补对科学的忽视。
Share this article
COPIED!
Newsletter
Get Quanta Magazine delivered to your inbox
Monica Almeida for Quanta Magazine
That’s a dramatic thing to say, that science has abandoned cause and effect. Isn’t that exactly what all of science is about?
Of course, but you cannot see this noble aspiration in scientific equations. The language of algebra is symmetric: If X tells us about Y, then Y tells us about X. I’m talking about deterministic relationships. There’s no way to write in mathematics a simple fact — for example, that the upcoming storm causes the barometer to go down, and not the other way around.
Mathematics has not developed the asymmetric language required to capture our understanding that if X causes Y that does not mean that Y causes X. It sounds like a terrible thing to say against science, I know. If I were to say it to my mother, she’d slap me.
But science is more forgiving: Seeing that we lack a calculus for asymmetrical relations, science encourages us to create one. And this is where mathematics comes in. It turned out to be a great thrill for me to see that a simple calculus of causation solves problems that the greatest statisticians of our time deemed to be ill-defined or unsolvable. And all this with the ease and fun of finding a proof in high-school geometry.
说起来很戏剧化，科学已经抛弃了因果关系。这不正是所有科学的意义所在吗？
当然，但你无法在科学方程式中看到这种崇高的愿望。代数的语言是对称的：如果X告诉我们关于Y，然后ÿ告诉我们关于X。我说的是确定性关系。没有办法在数学中写出一个简单的事实——例如，即将到来的风暴会导致气压计下降，而不是相反。
数学还没有开发出捕捉我们理解的非对称语言，即如果X导致Y并不意味着Y导致X。我知道，说反对科学的话听起来很糟糕。如果我告诉我妈妈，她会打我一巴掌。
但科学更宽容：看到我们缺乏对不对称关系的微积分，科学鼓励我们创造一个。这就是数学的用武之地。当我看到一个简单的因果关系演算可以解决我们这个时代最伟大的统计学家认为定义不明确或无法解决的问题时，我感到非常兴奋。所有这一切都伴随着在高中几何中找到证明的轻松和乐趣。
You made your name in AI a few decades ago by teaching machines how to reason probabilistically. Explain what was going on in AI at the time.
The problems that emerged in the early 1980s were of a predictive or diagnostic nature. A doctor looks at a bunch of symptoms from a patient and wants to come up with the probability that the patient has malaria or some other disease. We wanted automatic systems, expert systems, to be able to replace the professional — whether a doctor, or an explorer for minerals, or some other kind of paid expert. So at that point I came up with the idea of doing it probabilistically.
Unfortunately, standard probability calculations required exponential space and exponential time. I came up with a scheme called Bayesian networks that required polynomial time and was also quite transparent.
几十年前，您通过教机器如何进行概率推理而在 AI 领域一举成名。解释当时人工智能领域发生了什么。
1980 年代初期出现的问题具有预测性或诊断性。医生查看患者的一系列症状，并想得出患者患有疟疾或其他疾病的可能性。我们希望自动化系统、专家系统能够取代专业人士——无论是医生、矿物勘探者，还是其他类型的付费专家。所以在那一点上，我想出了用概率来做这件事的想法。
不幸的是，标准概率计算需要指数空间和指数时间。我想出了一个叫做贝叶斯网络的方案，它需要多项式时间并且非常透明。
Yet in your new book you describe yourself as an apostate in the AI community today. In what sense?
In the sense that as soon as we developed tools that enabled machines to reason with uncertainty, I left the arena to pursue a more challenging task: reasoning with cause and effect. Many of my AI colleagues are still occupied with uncertainty. There are circles of research that continue to work on diagnosis without worrying about the causal aspects of the problem. All they want is to predict well and to diagnose well.
I can give you an example. All the machine-learning work that we see today is conducted in diagnostic mode — say, labeling objects as “cat” or “tiger.” They don’t care about intervention; they just want to recognize an object and to predict how it’s going to evolve in time.
I felt an apostate when I developed powerful tools for prediction and diagnosis knowing already that this is merely the tip of human intelligence. If we want machines to reason about interventions (“What if we ban cigarettes?”) and introspection (“What if I had finished high school?”), we must invoke causal models. Associations are not enough — and this is a mathematical fact, not opinion.
然而，在您的新书中，您将自己描述为当今 AI 社区中的叛教者。凭什么？
从某种意义上说，一旦我们开发出使机器能够进行不确定性推理的工具，我就离开了舞台去追求更具挑战性的任务：因果推理。我的许多 AI 同事仍然被不确定性所困扰。有一些研究圈子继续致力于诊断，而不必担心问题的因果关系。他们想要的只是预测好，诊断好。
我可以给你举个例子。我们今天看到的所有机器学习工作都是在诊断模式下进行的——比如，将物体标记为“猫”或“老虎”。他们不在乎干预；他们只是想识别一个物体并预测它会如何随着时间的推移而演变。
当我开发出强大的预测和诊断工具时，我感到一种叛教，我已经知道这只是人类智能的一角。如果我们想让机器对干预（“如果我们禁止吸烟会怎样？”）和内省（“如果我高中毕业会怎样？”）进行推理，我们必须调用因果模型。关联是不够的——这是一个数学事实，而不是观点。
People are excited about the possibilities for AI. You’re not?
As much as I look into what’s being done with deep learning, I see they’re all stuck there on the level of associations. Curve fitting. That sounds like sacrilege, to say that all the impressive achievements of deep learning amount to just fitting a curve to data. From the point of view of the mathematical hierarchy, no matter how skillfully you manipulate the data and what you read into the data when you manipulate it, it’s still a curve-fitting exercise, albeit complex and nontrivial.
人们对人工智能的可能性感到兴奋。你不？
尽管我研究了深度学习正在做什么，但我发现它们都停留在关联级别上。曲线拟合。这听起来像是亵渎神灵，说深度学习的所有令人印象深刻的成就都只是为数据拟合了一条曲线。从数学层次的角度来看，无论您如何巧妙地操作数据以及在操作数据时读入数据的内容，它仍然是曲线拟合练习，尽管复杂且不平凡。
Monica Almeida for Quanta Magazine
The way you talk about curve fitting, it sounds like you’re not very impressed with machine learning.
No, I’m very impressed, because we did not expect that so many problems could be solved by pure curve fitting. It turns out they can. But I’m asking about the future — what next? Can you have a robot scientist that would plan an experiment and find new answers to pending scientific questions? That’s the next step. We also want to conduct some communication with a machine that is meaningful, and meaningful means matching our intuition. If you deprive the robot of your intuition about cause and effect, you’re never going to communicate meaningfully. Robots could not say “I should have done better,” as you and I do. And we thus lose an important channel of communication.
你谈论曲线拟合的方式，听起来你对机器学习印象不深。
不，我印象非常深刻，因为我们没想到通过纯曲线拟合可以解决这么多问题。事实证明他们可以。但我问的是未来——下一步是什么？你能有一个机器人科学家来计划一项实验并为悬而未决的科学问题找到新的答案吗？这是下一步。我们还想与一台有意义的机器进行一些交流，而有意义的意味着符合我们的直觉。如果你剥夺了机器人对因果关系的直觉，你将永远无法进行有意义的交流。机器人不能像你我一样说“我应该做得更好”。我们因此失去了一个重要的沟通渠道。
What are the prospects for having machines that share our intuition about cause and effect?
We have to equip machines with a model of the environment. If a machine does not have a model of reality, you cannot expect the machine to behave intelligently in that reality. The first step, one that will take place in maybe 10 years, is that conceptual models of reality will be programmed by humans.
The next step will be that machines will postulate such models on their own and will verify and refine them based on empirical evidence. That is what happened to science; we started with a geocentric model, with circles and epicycles, and ended up with a heliocentric model with its ellipses.
Robots, too, will communicate with each other and will translate this hypothetical world, this wild world, of metaphorical models.** **
拥有与我们分享因果直觉的机器的前景如何？
我们必须为机器配备环境模型。如果机器没有现实模型，你就不能指望机器在现实中智能地运行。第一步，可能在 10 年内发生，是现实的概念模型将由人类编程。
下一步将是机器将自行假设此类模型，并根据经验证据对其进行验证和完善。这就是科学发生的事情；我们从一个带有圆和本轮的地心模型开始，最后是一个带有椭圆的日心模型。
机器人也将相互交流，并将通过隐喻模型来翻译这个假设的世界，这个狂野的世界。** **
When you share these ideas with people working in AI today, how do they react?
AI is currently split. First, there are those who are intoxicated by the success of machine learning and deep learning and neural nets. They don’t understand what I’m talking about. They want to continue to fit curves. But when you talk to people who have done any work in AI outside statistical learning, they get it immediately. I have read several papers written in the past two months about the limitations of machine learning.
当您与当今从事 AI 工作的人分享这些想法时，他们有何反应？
AI目前是分裂的。首先，有些人陶醉于机器学习、深度学习和神经网络的成功。他们不明白我在说什么。他们想继续拟合曲线。但是当你与在统计学习之外在人工智能领域做过任何工作的人交谈时，他们会立即明白。我读了过去两个月写的几篇关于机器学习局限性的论文。
Are you suggesting there’s a trend developing away from machine learning?
Not a trend, but a serious soul-searching effort that involves asking: Where are we going? What’s the next step?
That was the last thing I wanted to ask you.
I’m glad you didn’t ask me about free will.
In that case, what do you think about free will?
We’re going to have robots with free will, absolutely. We have to understand how to program them and what we gain out of it. For some reason, evolution has found this sensation of free will to be computationally desirable.
In what way?
You have the sensation of free will; evolution has equipped us with this sensation. Evidently, it serves some computational function.
Will it be obvious when robots have free will?
I think the first evidence will be if robots start communicating with each other counterfactually, like “You should have done better.” If a team of robots playing soccer starts to communicate in this language, then we’ll know that they have a sensation of free will. “You should have passed me the ball — I was waiting for you and you didn’t!” “You should have” means you could have controlled whatever urges made you do what you did, and you didn’t. So the first sign will be communication; the next will be better soccer.
RELATED:
-
Concerns of an Artificial Intelligence Pioneer
-
Why Artificial Intelligence Like AlphaZero Has Trouble With the Real World
-
Artificial Intelligence Learns to Learn Entirely on Its Own
Now that you’ve brought up free will, I guess I should ask you about the capacity for evil, which we generally think of as being contingent upon an ability to make choices. What is evil?
It’s the belief that your greed or grievance supersedes all standard norms of society. For example, a person has something akin to a software module that says “You are hungry, therefore you have permission to act to satisfy your greed or grievance.” But you have other software modules that instruct you to follow the standard laws of society. One of them is called compassion. When you elevate your grievance above those universal norms of society, that’s evil.
So how will we know when AI is capable of committing evil?
When it is obvious for us that there are software components that the robot ignores, consistently ignores. When it appears that the robot follows the advice of some software components and not others, when the robot ignores the advice of other components that are maintaining norms of behavior that have been programmed into them or are expected to be there on the basis of past learning. And the robot stops following them.
This article was reprinted on TheAtlantic.com.
您是在暗示有一种远离机器学习的趋势吗？
不是一种趋势，而是一种严肃的反省努力，包括问：我们要去哪里？下一步是什么？
这是我想问你的最后一件事。
我很高兴你没有问我关于自由意志的问题。
在这种情况下，您如何看待自由意志？
我们将拥有自由意志的机器人，绝对的。我们必须了解如何对它们进行编程以及从中获得什么。出于某种原因，进化发现这种自由意志的感觉在计算上是可取的。
以什么方式？
你有自由意志的感觉；进化为我们提供了这种感觉。显然，它提供了一些计算功能。
机器人有自由意志时会很明显吗？
我认为第一个证据是机器人是否开始反事实地相互交流，比如“你应该做得更好”。如果一队踢足球的机器人开始用这种语言交流，那么我们就会知道他们有一种自由意志的感觉。“你应该把球传给我——我在等你，你没有！” “你应该有”意味着你可以控制任何促使你做你所做的事情的冲动，而你没有。所以第一个迹象将是沟通；下一个将是更好的足球。
有关的：
既然你已经提出了自由意志，我想我应该问你邪恶的能力，我们通常认为它取决于做出选择的能力。什么是邪恶？
这是一种信念，即您的贪婪或不满会取代社会的所有标准规范。例如，一个人有一个类似于软件模块的东西，上面写着“你饿了，因此你有权采取行动来满足你的贪婪或不满。” 但是您还有其他软件模块可以指导您遵守社会标准。其中之一叫做慈悲。当你将你的不满提升到社会的普遍规范之上时，那是邪恶的。
那么我们如何知道人工智能何时能够作恶呢？
当我们很明显有机器人忽略的软件组件时，就会一直忽略。当机器人似乎遵循某些软件组件而不是其他软件组件的建议时，当机器人忽略其他组件的建议时. 机器人停止跟随他们。
本文转载于TheAtlantic.com。
